Time-adaptive pattern recognition and prediction for situational robustness

In automatic target recognition (ATR) and time-series prediction, we use a training or historical database to tune the parameters of a classifier, which then operates on unknown data. In general, statistical properties in terms of class-conditional probability density functions are assumed to be similar between training and test data. Unfortunately, in many real-world applications, this assumption is violated, leading to erratic or even poor ATR performance. In this paper, we explore several methodologies to overcome this problem. We assess the efficacy of two algorithms with real data to demonstrate the importance of simple, intuitive solutions in dealing with difficult situations.